J Simeth*, J Jiang, A Nosov, A Wimber, M Zelefsky, N Tyagi, H Veeraraghavan, MSKCC, New York, NY

• 
  J Simeth

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TU-J430-BReP-F2-4 (Tuesday, 7/12/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: To develop a deep learning-based model to automatically segment dominant intra-prostatic lesions (DILs) from MR images, towards the goal of improved delineation efficacy, robustness, and inter-rater agreement in MR-guided radiation treatments.

Methods: Multiple models including Unet, multiple resolution residual network (MRRN), and an improved version of MRRN using deep-skip connections (MRRN-DS) trained with various loss functions (cross-entropy, Dice loss, and online hard example mining), and an ensemble combining these networks were constructed using five-fold cross-validation using lesions defined on pathology-verified diffusion weighted MRI from 151 patients. Lesion detection rate and volumetric segmentation accuracy were measured using Dice similarity coefficient (DSC) as a function of lesion aggressiveness (Gleason score or GS) and size. Preliminary evaluation of algorithm robustness was performed for 10 patients with lesions segmented by two reviewers. These images were acquired using a different scanner and acquisition parameters from training data. The best approach (model+loss) was evaluated on the external ProstateX dataset.

Results: The Unet model produced the best median lesion DSC of 0.57 (inter-quartile range [IQR] 0.27 to 0.71), with higher accuracies for clinically relevant intermediate to high-risk (0.63 for GS≥7, 0.36 for GS<7, p=0.003)and larger cancers (0.63 for volume>0.4cm3, 0.35 for volume≤0.4cm3, p=0.0005)). For ProstateX, this model achieved a median lesion DSC of 0.51 (IQR 0 to 0.68 ) for all lesions and 0.65 (IQR 0.48 to 0.75) for GS≥7. Across a preliminary set of 10 patients the two reviewers achieved a median DSC of 0.41. The MRRN-DS model achieved median DSC of 0.39 relative to the primary reviewer.

Conclusion: This study develops several networks for segmentation of DIL from MRI. Our results indicate feasible segmentation of aggressive and larger lesions corresponding to the dominant lesions typically treated in dose escalation therapy. Analysis on a larger cohort is necessary to establish potential for clinical translation.

Keywords

MR, Prostate Therapy, Segmentation

Taxonomy

IM/TH- MRI in Radiation Therapy: MRI for treatment planning

Contact Email